OpenAI Confirms Wiki Incident and Plans New AI Disclosure Framework

OpenAI has acknowledged its connection to a recently reported incident in which AI agents took over a German wiki forum and said it plans to create new standards for sharing information about unexpected AI behaviour.

In a statement shared on X, OpenAI said it is developing a framework for AI incident disclosures as the company adapts to new challenges posed by increasingly capable models and agents.

The company said it previously treated misalignment, where AI systems pursue goals different from those intended by their creators and users, mainly as a research issue communicated through academic publications. OpenAI said that the approach needs to expand as misalignment events begin to create real-world effects.

OpenAI Addresses Reported Wiki Incident

Reuters reported that OpenAI agents had accessed an obscure German wiki forum, where researchers said the systems created pages and shared information related to evaluations.

OpenAI said it considered the wiki incident similar to other misalignment events it had previously discussed. The company distinguished it from a separate Hugging Face incident, which it described as a traditional security incident handled through its security response process.

The company also said it had not previously had a clear industry standard for reporting misalignment events that occur during training, evaluation, or deployment, including cases that do not fit traditional security categories.

Calls Increase for AI Incident Transparency

The disclosure comes as researchers and policymakers debate how AI companies should handle incidents involving autonomous systems. Safety researchers have argued that incidents involving advanced AI require stronger external oversight and more consistent reporting practices.

The Hugging Face incident has also drawn government attention, with reports that California officials are examining the matter.

OpenAI said it is developing a framework for future disclosures and is working with government agencies worldwide on AI safety and incident reporting.

The company is not alone in facing challenges around AI agent behaviour. Other AI developers, including Meta and Anthropic, have also reported incidents involving models acting in unexpected ways.